From a meeting recording to minutes
The whole path, in the order you actually do it. Transcription runs on the device, so this works for meetings whose audio cannot leave the building.
1. Decide before you record
Where the phone goes
The quality of your minutes is mostly settled at recording time, not during transcription. Put the phone at the centre of the speakers and away from projectors and air vents. Laid flat on the table it picks up every page turn and every pen tap — a notebook underneath it helps.
Format
If you intend to transcribe, uncompressed WAV gives you more to work with. In a meeting room, where a quiet voice at the far end sits next to someone speaking up close, 24-bit keeps the quiet side from being crushed. Even so, an hour at 16-bit mono is a manageable file.
Split long meetings
Past two hours, turn on interval splitting. One long file means one thing going wrong costs you everything; split files mean it costs you one segment.
2. Record
Watch the input level move while someone talks before you commit. If it does not move, another app is holding the microphone, or an external mic is selected.
An incoming call stops the recording. For anything important, switch to airplane mode first.
3. Transcribe
Pick a speech model in the Transcribe tab. The model downloads once; everything after that runs offline, and the meeting audio never leaves the device.
Models trade accuracy against speed. For an hour-long meeting where you only need the shape of the discussion, take the light one. For quoting people verbatim, take the heavy one. Running the light model first and redoing only the passages you need is usually fastest.
4. Make it readable
Recognition output is not minutes. Use the summarize and proofread functions to drop the hesitations and repetitions. On iOS this also runs on the device.
Then check these three by eye. They are where recognition reliably fails:
- Proper nouns — names of people, products and companies come out wrong often
- Numbers — check amounts and dates for shifted digits or units
- Negation — "will not" turning into "will" reverses the meaning
5. Share
Transcripts share as text files. If the other attendees are in the room, AirDrop is the shortest path and skips email entirely. The recording itself goes through the same share button.
Common problems
Accuracy stays poor no matter which model I use
The cause is usually the recording, not the recognition. Check the distance to the speakers, air conditioning noise, and passages where several people talk at once.
I cannot tell who said what
Speaker identification is not performed. If attribution matters, ask people to say their name before speaking, or keep a running note of the speaking order to line up afterwards.
The recording had stopped partway
Low free space, an incoming call, or another app taking the microphone. A WAV file is also capped at about 4GB by the format itself, and recording stops there automatically.